This study developed and validated machine learning (ML) models for screening anemia and leukemia using complete blood count (CBC) features. Data were collected from 287 CBC reports, with 19 hematologic and 2 demographic features. To address small sample size and class imbalance, the authors generated 2,000 synthetic cases and combined them with original data to create a hybrid dataset (n=2,287). Four diagnostic classes were considered: normal, anemia, leukemia, and anemia+leukemia (“combination”). Feature selection by point-biserial correlation and recursive feature elimination identified a “fingerprint” of 14 CBC parameters with statistical and clinical significance. Six ML algorithms were tested (decision tree, random forest, SVM, logistic regression, gradient boosting, multilayer perceptron). Random forest performed best, achieving 98% accuracy, with macro-averaged precision 97%, recall 98%, specificity 99%, and miss-rate 2%. Gradient boosting achieved comparable results (97% accuracy). External validation on an independent dataset of 270 CBC reports reduced accuracy to 74%, with recall highest for leukemia (89%) and combination (95%) but weaker for normal (78%) and anemia (63%).